Nanyang Technological University Singapore’s Centre of AI in Medicine is exploring how healthcare organisations can move beyond automation by building what it describes as an “AI brain”: systems that combine internal data with external signals to support forward-looking decision making. Its Chief Data Scientist Dr Wilson Goh said this requires rethinking infrastructure, data access, and governance.
In this interview, Goh discusses how healthcare organisations can build these capabilities, the types of data and computing environments required, and where AI use cases are most likely to scale beyond the pilot stage.
How can organisations build an “AI brain” for strategic decision making?
Building an “AI brain” means developing systems that go beyond automation to continuously learn, reason, and adapt within an organisation’s data ecosystem. This requires connecting internal data with external signals — such as market movements, patient trends, or supply chain dynamics — to generate forward-looking, strategic insights.
With the rise of agentic models, we are seeing a shift towards decision-driven architectures, where AI systems are integrated with enterprise knowledge graphs and scenario-planning frameworks, allowing them to act with a degree of agency rather than remaining passive analytical tools.
That said, we need to be mindful of how much autonomy and decision-making authority we grant these systems. Automation must always be balanced with considerations of risk, safety, and human development. Over-automation could deprive younger professionals of the experiential learning needed for expertise and judgment.
In the pursuit of efficiency and optimisation, we must not lose sight of humanism, ensuring that technology augments, rather than erodes, the growth and agency of people.
What IT infrastructure and LLMs support your AI use cases?
At NTU, we operate our own high-performance computing servers to support a wide range of research activities. For modelling work that involves healthcare and national datasets, our Centre of AI in Medicine is the only unit on campus with a dedicated microaccess lab that provides secure access to the Ministry of Health’s Trust and Healix systems. These environments function as integrated trusted research environments, enabling controlled data access and model development within regulatory safeguards.
In terms of LLMs, we adopt a flexible approach depending on data sensitivity. For operational or non-clinical applications — such as chatbots for general administrative support — we use commercially available models from providers like Anthropic and OpenAI. However, when the work involves patient data or other sensitive information, we turn to more tightly controlled ecosystems and locally adaptable models such as Qwen or LLaMA to ensure compliance with privacy and governance requirements.
What data are you using for your AI use cases, and how are you using them?
We work with a range of data sources, including sociodemographic and psychometric data, video and speech recordings, electronic health records, digital wearable data, and high-throughput biomolecular platforms such as genomics, proteomics, and metabolomics.
These datasets are used for feature engineering and multimodal model development, as well as for fine-tuning LLMs to support diagnosis, prognosis, clinical decision-making, and personalised health management.
Which AI use cases in healthcare are most scalable, and how can organisations move beyond pilots?
The AI use cases most likely to scale rapidly are those with lower risk yet clear and demonstrable impact. Clinical triage systems are a prime example: AI tools that can quickly determine whether patients should go home, be discharged, or be admitted offer immediate time, cost, and resource savings for both hospitals and patients.
Drug discovery pipelines are another proven domain for acceleration, with AI helping to shorten clinical testing workflows and improve the precision of candidate selection.
Equally promising is drug repurposing, where AI uses existing pharmacological and safety data to identify new therapeutic uses for approved drugs. This approach not only gives compounds a “second life” but also enables the management of comorbidities, helping one drug address multiple related conditions in a single patient.
Another scalable area is hospital operations optimisation. Though less glamorous, it represents low-risk, high-efficiency gains through resource planning and predictive maintenance.
Beyond these, I believe another area deserves significant attention: AI in patient risk profiling and trajectory stratification, particularly for chronic disease management. As societies age and comorbidities rise, scalable solutions for continuous monitoring and stratification will be essential to prevent healthcare systems from becoming overwhelmed.
Moving beyond the pilot phase involves both technical and human challenges. Before initiating any pilot, there must be robust reviews to ensure technical soundness, workflow integration, and stakeholder alignment. Early economic and operational assessments are crucial to avoid promising technologies failing right after proof of concept. Adoption barriers often stem from limited trust or familiarity, rather than algorithmic shortcomings.
These are the conversations that are increasingly discussed across the industry, including at platforms such as Asia Tech x Enterprise. Policymakers, alongside healthcare leaders and technology practitioners, exchange perspectives on deploying AI responsibly and at scale in real-world environments, helping to surface the practical considerations needed to move beyond isolated pilots.
Successful implementation requires interdisciplinary teams that bring together data scientists, clinicians, and social scientists to coordinate across departments and facilitate dialogue from leadership to ground-level execution.
Above all, organisations must cultivate a culture of openness and technological agility. Even the most effective AI system can fail if people are unwilling or unprepared to adopt it.
How can healthcare organisations implement AI governance without slowing innovation?
The key is to co-design governance alongside innovation. Regulatory principles such as transparency, fairness, and explainability should be embedded directly into model design and evaluation pipelines.
At the NTU Centre of AI in Medicine, governance is built into the development process. As we develop models, we run parallel validation frameworks and simulated enactments to assess both predictive performance and ethical compliance in tandem.
Frequent and open engagement with experts, stakeholders, and the public is essential. This shifts governance from a box-ticking compliance exercise into a collaborative, participatory design process. When governance evolves in step with innovation, it prevents costly rework later and strengthens trust, supporting responsible adoption across healthcare ecosystems.
